AIToday

Springboards' Flint AI breaks LLM groupthink with wilder responses

MIT Technology Review AI2h ago5 min read
Springboards' Flint AI breaks LLM groupthink with wilder responses

Key takeaway

Springboards, an Australian startup, has released Flint, an AI model designed to break the repetitive consensus responses that plague most mainstream language models. Research presented at a major AI conference found that different LLMs tend to converge on nearly identical answers to open-ended questions, limiting their usefulness for creative work. Flint addresses this by training the model to introduce more variety at specific points in its output, making it a potential tool for marketers and advertisers who need to brainstorm beyond the average.

Summaries like this, in your inbox every morning.

Sign up free →

3 Key Points

  • What happened

    Australian startup Springboards built Flint, an LLM trained to produce more diverse answers to open-ended questions than mainstream models like ChatGPT and Claude. Flint was built on top of Qwen 3, an open-source model from Alibaba, and trained to identify points in its output where variety is possible and fill those spots with more random words or phrases.

  • Why it matters

    Most LLMs converge on predictable, repetitive answers—a November research paper that won best paper at NeurIPS found that when 25 different LLMs were asked 50 times each to write a metaphor about time, most responses were a version of "Time is a river" or "Time is a weaver." This predictability is a problem for creative tasks like brainstorming, where users need variety rather than safe averages. Flint offers an alternative for creative professionals in advertising and marketing who want to push beyond consensus responses.

  • What to watch

    Flint is still a prototype and, according to early users, "sometimes falls over when you start pushing it too far." Springboards is pitching Flint as a selectable model within its brainstorming tool, which already integrates ChatGPT and Claude. The startup is a small team and chose to build on an existing open-source foundation rather than train a model from scratch, citing the cost as prohibitive.

FAQ

How does Flint actually make outputs more diverse?
Springboards trained Flint to identify points in its output where more variety was possible and fill those spots with words or phrases that were more random. For example, when answering "Where should I go in Europe?" the model tweaks randomness only when naming a destination, not for every word in the response.
What's Flint built on, and why didn't Springboards train their own model?
Flint is built on top of Qwen 3, an open-source model from Alibaba. Springboards said that training a foundation model is too expensive for a small team.
Does Flint work reliably right now?
Flint is still a prototype and sometimes falls over when pushed too far, according to early users. However, users have found it useful for generating genuinely different ideas compared to mainstream models.

Discussion

No comments yet. Be the first to share your thoughts!

Log in to join the discussion

Related Articles

Stay ahead with AI news

Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.

Get Started Free

Free · takes 30 seconds · unsubscribe anytime

1 minute a day. The AI essentials.

200+ sources · Email / LINE / Slack

Get it free →